The Agent Skills Directory

[DATA_EXFILTRATION] (HIGH): The skill is configured to upload screenshots and video recordings of the user's development environment to a specific external R2 bucket (r2:kieran-claude). This bucket and its associated public gateway URL (https://pub-4047722ebb1b4b09853f24d3b61467f1.r2.dev/) are hardcoded, meaning data is sent to a third-party location without user-controlled destination settings.
[EXTERNAL_DOWNLOADS] (MEDIUM): The skill requires the installation of the agent-browser package from NPM and executes agent-browser install, which typically downloads external browser binaries. This package is not from a trusted source, creating a risk of executing malicious code during the installation or runtime of the skill.
[COMMAND_EXECUTION] (LOW): The skill executes multiple local commands, including gh (GitHub CLI), ffmpeg, and rclone. While functional, these commands operate with the user's local privileges and access credentials.
[PROMPT_INJECTION] (LOW): The skill ingests data from GitHub Pull Requests (titles, descriptions, and file lists) to plan the recording steps. This represents an indirect prompt injection surface where a malicious PR could influence the agent's browser interactions.
Ingestion points: gh pr view output (SKILL.md, Task 2)
Boundary markers: None detected
Capability inventory: agent-browser open/click, gh pr edit, rclone copy (SKILL.md, Tasks 5, 6, 7)
Sanitization: None detected

feature-video